The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface within the scaffold_agentic_workflow.py script. It generates .agent.md and .yml files by directly interpolating the 'description' and 'body' of a source skill into the generated instructions and prompts.
Ingestion points: The scaffold_agentic_workflow.py script reads the content of an existing SKILL.md file provided via the --skill-dir or --plugin-dir arguments.
Boundary markers: The generated .agent.md file uses YAML frontmatter delimiters (---) for the description but does not employ protective delimiters or "ignore embedded instructions" warnings for the main instruction body.
Capability inventory: The generated GitHub Actions workflows grant the agent the ability to use the @github/copilot CLI with powerful tools including read, write, and shell.
Sanitization: The script lacks sanitization for the instruction content being copied, which could allow a malicious source skill to embed instructions that are executed by the generated agent in the CI/CD environment.
[COMMAND_EXECUTION]: Several utility scripts in the scripts/ directory utilize the subprocess module to execute system commands and interact with CLI tools.
improve_description.py and run_eval.py execute claude -p or copilot -p commands.
generate_review.py executes the lsof command to manage local server ports and uses os.kill to terminate existing processes.
[EXTERNAL_DOWNLOADS]: The generated GitHub Actions runner file includes a step to install the @github/copilot package globally via npm. This is a well-known package from a trusted organization (GitHub). Additionally, documentation patterns like dynamic-specification-fetching.md suggest fetching SDK references from official GitHub repositories.

create-agentic-workflow